MADReg: Modality-aware deformation regularization for unsupervised infrared-visible image registration

Infrared-visible image registration establishes spatial correspondence for multisensor fusion and perception. In unsupervised deformation learning, differences in modality responses can be mistaken for geometric misalignment, causing erroneous displacement and distortion of scene geometry. Common deformation priors encourage spatial regularity, invertibility, or topology preservation, whereas determining whether a deformation is supported by the observed geometry requires additional guidance. We propose MADReg, a modality-aware deformation regularization framework that obtains complementary geometric references from reverse registration and translation-assisted registration. Bidirectional Deformation Field Consistency (BDFC) cross-utilizes the forward and backward fields and constrains the structural correspondence of the resulting warped images. Translation-Driven Deformation Regularization (TDDR) estimates auxiliary fields from translated same-modality pairs and uses their spatial displacements and amplitude and phase spectra to guide the main field. Together, these reference processes relate the estimated deformation to shared scene structures and the overall deformation patterns of auxiliary mappings, helping suppress spatial changes induced by modality responses. Within-dataset and direct cross-dataset experiments on Visnir and RoadScene demonstrate improved correspondence accuracy relative to representative methods. Ablation studies and visualizations support the complementary value of the geometric references for preserving scene structure and controlling deformation irregularity.